Search

CN-122017779-A - HRRP target identification method and system based on angle guidance

CN122017779ACN 122017779 ACN122017779 ACN 122017779ACN-122017779-A

Abstract

The invention discloses an HRRP target recognition method and system based on angle guidance, comprising the steps of constructing a target recognition network model, acquiring HRRP data of a target vehicle in a static state, and carrying out dynamic convolution operation on the HRRP data by introducing azimuth information so as to extract angle self-adaptive local scattering characteristics; and finally, the local dynamic convolution characteristic and the global attention characteristic are subjected to cross fusion, and the target category is output through the classification module. According to the invention, the angle condition kernel mixed convolution module is used for introducing azimuth information into a convolution kernel generation process, so that the convolution kernel can be adaptively adjusted according to the azimuth, the change rule of HRRP along with the azimuth is effectively described, the traditional azimuth invariant feature extraction or angular domain division is not required, the influence of azimuth sensitivity on the identification performance is relieved, and the identification precision and generalization capability are further improved.

Inventors

  • HUI HUI
  • LI WEI
  • ZHU RUIXUAN
  • YAO JUNLIANG

Assignees

  • 西安理工大学

Dates

Publication Date
20260512
Application Date
20260122

Claims (9)

  1. 1. The HRRP target identification method based on angle guidance is characterized by comprising the steps of constructing a target identification network model, carrying out dynamic convolution operation on HRRP data by introducing azimuth angle information by using HRRP echo data in a stationary state of a target vehicle in a data set, extracting local scattering characteristics of angle adaptation, further obtaining global structural characteristics by using a global attention module, and finally carrying out cross fusion on the local dynamic convolution characteristics and the global attention characteristics and outputting target types by a classification module.
  2. 2. The HRRP target identification method based on angle guidance of claim 1, which specifically comprises the steps of: Step 1, constructing a target recognition network model, wherein the target recognition network model comprises a preprocessing module, an angle condition kernel mixed convolution module, a double-branch feature extraction module, a feature fusion module and a classification recognition module; step 2, acquiring HRRP data of the target vehicle in a stationary state, and constructing a high-resolution range profile HRRP data set =( , ) I=1, 2,.. P, P denotes the number of samples, Representing azimuth angle , Represents HRRP distance; step 3, carrying out normalization and center of gravity alignment processing on the HRRP distance, and carrying out two-dimensional coding processing on the azimuth angle; Step 4, inputting the azimuth angle processed in the step 3 into an angle condition kernel mixed convolution module, mapping the azimuth angle into Gaussian distribution with variable mean and variance, and generating dynamic convolution kernel parameters by using a re-parameterization method; Step 5, using the dynamic convolution kernel parameters generated in the step 4 to extract local features in the double-branch feature extraction module, and extracting global features from global branches in the double-branch feature extraction module; step 6, fusing global features and local features by using a cross attention mechanism of a feature fusion module, taking the global features as inquiry, taking the local features as keys and values, and guiding a network to selectively pay attention to the features which are most critical to discrimination so as to obtain fused features; step 7, inputting the fusion characteristics into the classification module to obtain classification results; and 8, setting classification loss according to the classification result, training the model by using the fused feature set center loss, and identifying the vehicle HRRP to be identified by using the trained target identification model.
  3. 3. The method and system for identifying an HRRP target based on angle guidance according to claim 2, wherein step 3 is specifically as follows: Step 3.1, carrying out normalization and barycenter alignment treatment on each HRRP sample, wherein the normalized samples are as follows: Wherein, the method comprises the steps of, The normalized sample is represented by a representation of the sample, Representing the calculated 2-norm; each HRRP distance Comprising n distance units of the total number of the distance units, , ; Representing the p-th distance unit in the i-th sample; The center of gravity is aligned and processed, Wherein, C represents the center of gravity of the HRRP sample, n represents the number of distance units, and then the center of each HRRP sample is translated to the center of gravity, namely, the center of gravity alignment operation is completed; Step 3.2, for the i-th sample azimuth Performing radian conversion, calculating sine value and cosine value, and generating azimuth angle coding vectors of all samples ; Wherein, the For the azimuth angle, Is the processed azimuth angle.
  4. 4. The HRRP target identification method and system based on angle guidance of claim 3, wherein the generating process of generating the dynamic convolution kernel parameters in the step 4 is as follows: Step 4.1, the azimuth angle after pretreatment As condition input, the average value generating branch composed of three-level full-connection network generates an average value vector of output dynamic convolution kernel parameters, Wherein, the method comprises the steps of, And Respectively representing the weight and bias of the full connection layer, wherein the average value generation branch consists of a three-layer full connection network and two ReLU activations, and the average value vector The output dimension of (2) is preset to ; Wherein, the Representing the number of parallel dynamic convolution kernels, =8, The number of channels representing the input profile, The number of channels representing the output profile, Representing the size of a one-dimensional convolution kernel; And according to the number of the parallel dynamic convolution kernels Will be Is reconstructed into Wherein each set of mean parameters Corresponding to a set of candidate dynamic convolution kernel parameter values; Wherein each set of mean parameters Values corresponding to a set of candidate dynamic convolution kernel parameters Step 4.2, the azimuth angle after pretreatment As a conditional parallel input to a variance generating branch, a network packet of the variance generating branch Wherein, the The representation Softplus activates the function, The Sigmoid function is represented as a function, The variance generation branch is composed of three layers of fully connected networks, two Softplus activations and the last layer of Sigmoid activations for presetting variance constraint parameters; The variance vector Is also reconstructed as Each set of variance parameters The discrete degree is used for describing the corresponding dynamic convolution kernel parameters; Step 4.3, regarding the mean parameter and the variance parameter as the distribution parameters of the dynamic convolution kernel random variable, for the first Resampling and gradient updating are carried out on the group dynamic convolution kernel: Wherein, the Represent the first The dynamic convolution kernel parameters of the group, Representing a standard normal distribution; Step 4.4, for preprocessing the input samples Performing global average pooling and inputting into an attention network to obtain The weight of the weight is calculated, Wherein the attention network comprises a global average pooling layer, a flattening layer, a full connection layer, a ReLU activation layer, a full connection layer, a Sigmoid activation layer, Represent the first The group dynamic convolution checks the currently entered response weight, , , , Is the weight and parameters of the full connection layer; step 4.5, the The group dynamic convolution kernel is fused according to the attention weight to obtain a final convolution kernel , Step 4.6, performing dynamic convolution calculation, Wherein, the Representing the input features, y representing the output features, and C representing the dynamic convolution operation.
  5. 5. The HRRP target recognition method based on the angle supervision feature generation of claim 4, wherein in step 5, the dual-branch feature extraction module comprises a CNN local feature extraction branch composed of dynamic convolution and a global feature extraction branch composed of a transform encoder; the local feature extraction branch consists of three convolution block sequences and residual error connection, wherein the first one and the third one are static convolution blocks, and the second one is dynamic convolution kernel parameters generated by a conditional kernel mixed convolution module The dynamic convolution block introduces azimuth angle parameters in convolution operation to make the convolution weight possess azimuth angle perception characteristic, and the output from the module is local characteristic ; The global feature extraction module consists of a transducer encoder and comprises a multi-head self-attention layer, two fully-connected feedforward layers FFN and two LayerNorm layers, and the preprocessed sample is subjected to position coding, multi-head self-attention mechanism, layer normalization and feedforward network connection layers in sequence to obtain global feature representation aiming at the current sample 。
  6. 6. The HRRP target recognition method based on the angle supervision feature generation according to claim 1, wherein the feature fusion method in the step 6 is cross attention fusion, and comprises the following steps: step 6.1, a feature fusion module is constructed and divided into a multi-head cross attention mechanism and a feedforward network module FFN, wherein the cross attention can be expressed as: Wherein Q, K, V represents a query matrix, a key matrix and a value matrix respectively, which are obtained by performing 3 different linear transformations on input, and the dimension is The cross ideas in the cross attention mechanism are presented here in the introduction of the cross ideas into the expression Q, K, V, so that Q of the calculated attention is derived from the extracted global feature transform, while K and V are derived from the local feature transform: Wherein, the Is a matrix of weight values Q, K, V that can be learned, 、 Features extracted from different modules, cross-attention calculations are: Wherein, the The function of (1) is to transform the input vector into a high-dimensional space by performing linear transformation on the input, comprising 4 fully connected layers, 1 layer normalization layer, 1 Dropout layer and 1 Softmax operation; step 6.2, forming MHCA by stacking multiple cross-attentions, first performing multi-head division on Q, K, V in the second dimension to obtain h heads , Then, a cross-attention calculation is performed for each head: Splicing the output matrix of each head in the last dimension: Wherein, the Representing splicing operation, which consists of 3 linear transformation layers and a Softmax attention layer, and finally, outputting the signals through a characteristic fusion module as ;; Step 6.3 introducing a center penalty in the fused feature space of the cross-attention output ; The center loss is calculated by calculating the distance between the feature vector and the Euclidean distance square measure between the feature vector and the centers of the corresponding categories, and the expression is as follows: Wherein, the As a fusion feature of the i-th sample, For the class labels to which it corresponds, Is marked as Is defined in the center of the category of (c), Representing the L2 norm.
  7. 7. The method for identifying the HRRP target based on the angular supervision feature according to claim 1, wherein the classification module comprises a full connectivity layer and a softmax activation layer in step 7; The cross entropy classification loss used is: Wherein N is the number of categories, The real tag one-hot code is encoded, Predicting class probabilities for the model; The total loss function is: Wherein, the The trade-off parameter is expressed, taking 0.2.
  8. 8. The HRRP target recognition system based on the angle supervision characteristic generation is characterized in that the system comprises a target recognition network model, wherein the target recognition network model comprises the following components: The preprocessing module is used for preprocessing the samples in the data set and performing two-dimensional processing on azimuth angles to obtain preprocessed samples and two-dimensional angles; the angle condition kernel mixed convolution module is used for mapping the azimuth angle into Gaussian distribution with variable mean and variance through a coding network, and generating a dynamic convolution kernel by using a heavy parameterization method; The double-branch feature extraction module is used for extracting local features from CNN local features formed by dynamic convolution kernels from the angle condition kernel mixing convolution module and global features from a transducer encoder; The classification module is used for obtaining classification results; The loss function module is used for setting classification loss according to the classification result and setting center loss by utilizing the fused characteristics; the target recognition module is used for training the target recognition model according to the classification loss and the central loss, obtaining a trained target recognition model, and recognizing the HRRP of the vehicle to be recognized by using the trained target recognition model.
  9. 9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 7 when the computer program is executed.

Description

HRRP target identification method and system based on angle guidance Technical Field The invention relates to the technical field of radar target detection, in particular to an HRRP target recognition method based on angle guidance and an HRRP target recognition system based on angle guidance. Background Because the radar has a long acting distance and can work around the clock, the vehicle target recognition research driven by radar measurement data has important significance in traffic monitoring. The high-resolution range profile (High Resolution Range Profile, HRRP) is a one-dimensional echo feature obtained by a radar target under the condition of high range resolution, and essentially reflects the range distribution and energy distribution characteristics of each scattering center of the target in the radar sight direction. Because HRRP has the advantages of low data dimension, high acquisition speed, strong adaptability to complex environments and the like, the method has been widely applied to the field of radar automatic target recognition. At present, a radar-based vehicle target recognition method has advanced to some extent, but some key disadvantages still exist. Firstly, the echo characteristics of the vehicle target under different observation azimuth, speed and gesture conditions are obviously different, so that the recognition performance is reduced, secondly, the radar echo data has larger noise and high labeling cost, the existing method has stronger dependence on large-scale labeling data, and the generalization capability is limited. Meanwhile, as a typical radar echo signal of a vehicle, a High Resolution Range Profile (HRRP) is widely used for vehicle target recognition research due to its clear structure and high information density. Currently, HRRP target recognition methods can be roughly classified into two types, a conventional HRRP target recognition method and a HRRP target recognition method based on deep learning. The traditional HRRP target recognition method mainly comprises a pattern classification method based on characteristic engineering and a statistical inference method based on a probability map model. The method based on the probability density estimation is generally based on the assumption that distance units are mutually independent, a joint probability model of each distance unit of the HRRP is constructed, and the judgment of the target category is realized by combining probability distribution of each distance unit. With the development of deep learning technology, the HRRP target recognition method based on the deep neural network gradually becomes a research hotspot. The method realizes the end-to-end joint optimization of HRRP characteristic representation learning and classification decision by constructing a multi-layer nonlinear mapping network, and compared with the traditional method, the method can automatically learn the characteristic representation with discriminant through a back propagation mechanism, thereby obviously reducing the dependence on the design of manual priori characteristics. Although the method improves the identification performance of the HRRP targets to a certain extent, due to the inherent azimuth sensitivity characteristic of the HRRP data, the HRRP obtained by the same target under different observation azimuth or attitude conditions often has significant differences, which provides a serious challenge for stable identification under different azimuth conditions. In addition, the azimuth angle is used as important auxiliary priori information, so that the accuracy of target identification can be effectively improved in theory, but in practical application, a training sample covering a complete azimuth angle range is often difficult to obtain, and the generalization capability of the model under the condition of no azimuth angle is still limited. Therefore, how to fully utilize azimuth information under the condition of incomplete sample azimuth and improve the adaptability of the model to the azimuth change becomes a key problem to be solved in the field of HRRP target identification. Disclosure of Invention The invention aims to provide an HRRP target identification method based on angle guidance, which solves the problems that in the identification of mark of mine categories, when the observation azimuth angle of a vehicle target changes, the identification performance is obviously reduced and the robustness is insufficient in the existing high-resolution range image HRRP target identification method. It is also an object of the present invention to provide an HRRP target recognition system based on angular guidance. The first technical scheme adopted by the invention is that the HRRP target recognition method based on angle guidance aims at that the scattering characteristics of a vehicle target under different observation orientations can be obviously changed, so that the range unit echo amplitude distribution of the HRRP